Memory - Computational Explorations in Cognitive Neuroscience - page 300

studies, monkeys saw sequences of stimuli and had to
respond when the first stimulus in the sequence was re-
peated. If we label the stimuli with letters (they were ac-
tually pictures presented on a computer monitor), such
a sequence would be: ABCDA , where the second A
should trigger a response. Miller et al. (1996) found
that neurons in the inferior-temporal cortex (IT) only
represented the most recently presented stimulus. In-
terestingly, these IT neurons did exhibit a reliable dif-
ference in activation when the initial stimulus was re-
peated, which can be attributed to the weight-based
priming mechanism discussed previously (section 9.2).
Thus, performance in this task could conceivably be
mediated either by this weight-based memory, or by
activation-based memory.
To force the use of activation-based memories,
Miller et al. (1996) used an ABBA variant of the task,
where the intermediate stimuli could also have repeats.
They found that monkeys initially tended to respond
to the second internal repeat (i.e., the second B ), but
that with training they could learn to use the frontal
activation-based memory of the initial stimulus to re-
spond only when this stimulus repeated. Thus, this se-
ries of experiments nicely captures the differences be-
tween the posterior and frontal cortex. The posterior
cortex (e.g., IT) can do simple weight-based priming,
but not robust active maintenance, whereas the pre-
frontal cortex can do the robust active maintenance of
stimuli in the face of ongoing processing.
Active maintenance has also been localized to the
prefrontal cortex in humans using functional magnetic
resonance imaging (fMRI) as subjects performed a
working memory task (Cohen, Perlstein, Braver, Nys-
trom, Noll, Jonides, & Smith, 1997). Importantly, this
study was specifically able to show that the prefrontal
cortex was active during the maintenance period, not
just at the start and end of it, which corroborates the
findings of sustained frontal activation in the monkey
recordings. Much of the other evidence for working
memory in the prefrontal cortex comes from more com-
plex tasks, that will be discussed in the context of
higher-level cognition in chapter 11. In the next section,
we focus on evidence suggesting that the prefrontal cor-
tex can dynamically regulate between robust mainte-
nance and rapid updating based on task demands.

9.5.1 Dynamic Regulation of Active Maintenance

There is substantial evidence to suggest that the neuro-
modulatory substance dopamine (DA) plays an impor-
tant role in regulating the frontal active memory system
(O'Reilly et al., 1999a; Cohen et al., 1996). DA ag-
onists (i.e., drugs that mimic or enhance the effect of
dopamine) have been found to improve memory perfor-
mance in humans (Luciana, Depue, Arbisi, & Leon,
1992), and DA antagonists (i.e., drugs that block the
effects of DA) interfere with performance in delayed-
response tasks in monkeys (Sawaguchi & Goldman-
Rakic, 1991), and directly affect frontal neuronal ac-
tivity (Williams & Goldman-Rakic, 1995). DA in
the frontal cortex typically synapses in combination
with other inputs, enabling it to modulate the effi-
cacy of these inputs (Lewis, Hayes, Lund, & Oeth,
1992; Williams & Goldman-Rakic, 1993), and has
been shown electrophysiologically to potentiate both
afferent excitatory and inhibitory signals (Chiodo &
Berger, 1986; Penit-Soria, Audinat, & Crepel, 1987).
Finally, DA has played an important role in explaining
frontal deficits associated with schizophrenia (Cohen &
Servan-Schreiber, 1992).
On the basis of the preceding evidence, Cohen
et al. (1996) proposed that the midbrain nuclei that send
dopamine to the frontal cortex (i.e., the ventral tegmen-
tal area, VTA ), under control of descending cortical pro-
jections, enable the frontal cortex to actively regulate
the updating of its representations by controlling the re-
lease of dopamine in a strategic manner. Specifically,
they proposed that the afferent connections into the
frontal cortex from other brain systems are usually rela-
tively weak compared to stronger recurrent excitation,
but that DA enhances the strength of these afferents
at times when rapid updating is necessary. Thus, DA
serves as a dynamic gating mechanism, such that when
DA is firing, the “gate” into the active memory repre-
sentations is open (leading to rapid updating), but oth-
erwise is closed (leading to robust maintenance). Note
that this dynamic regulation of the relative strength of
the input versus the recurrent maintenance connections
is just the kind of mechanism that we explored in the
previous simulations by manually adjusting the weight
scaling parameter. Hochreiter and Schmidhuber (1997)